Behind Micron's Rise: The Survival Battle, Capital Game, and AI Rebirth of America's Lone Memory Player

36Kr Auto Digital Kr Future Consumption Intelligent Emergence Future City Launch Power on 36Kr Global 36Kr Research Institute Chaotic Tide TIDE 36Kr Enterprise Service Review 36Kr Finance Workplace bonus 36 Carbon Post-Wave Research Institute Dark Surge Waves Hard Kr Kr Rui Research Institute Media Brand

Enterprise Accounts Enterprise Service Review 36Kr Research Institute 36Kr Innovation Consulting Enterprise Services

Core Services City Window Government Services

Venture Capital Release LP Source Plan VClub VClub Investment Institution Database Investment Institution Job Referrals Investor Certification Investor Services

Event Search Project Recommendations I Want to Join City Cooperation Behind Micron's Rise: The Survival Battle, Capital Game and AI Rebirth of America's Memory "Only Child" Silicon Valley 101 · October 18, 2026 17:00 The Unique "American-East Asian Duality"

On September 30, Micron once again delivered a record-breaking financial report.

In the fourth quarter of fiscal year 2026, Micron's single-quarter revenue reached $54.23 billion, with an operating margin of 80.7%, and both revenue and net profit set new records. Strong demand brought by AI is making this memory chip company enter the most profitable moment in its history.

But looking back at Micron's nearly fifty-year history, its growth was far less smooth than today's performance suggests. In the brutal price cycles of the memory industry, many former giants exited one after another, while Micron survived losses and industry reshuffles again and again, ultimately becoming the "only child" of the American memory industry.

In this process, Micron gradually developed a special way of survival: backed by American capital markets, through a series of "small fish swallowing big fish" acquisitions and investments, it continuously extended its manufacturing operations into Asia. Today, the vast majority of this American memory giant's production capacity is in Asia, while the U.S. government is offering billions of dollars in subsidies, hoping to bring its advanced manufacturing capabilities back to the United States.

The starting point of all this, however, was in Idaho, far from Silicon Valley and known as America's "Potato State." In the basement of a dental clinic, Micron took its first step as a startup. And its ticket into the memory chip manufacturing industry came from a "potato tycoon" who made his fortune supplying frozen french fries to McDonald's.

From starting with investment from local businessmen, to expanding with the help of American capital and Asian manufacturing, and then to becoming a key company in the AI supply chain today, how exactly has Micron managed to survive industry downturns again and again? Can its nearly fifty years of survival experience help it win the next round of AI competition?

Although located in the "Potato State," Micron is not a company that sprang out of a crack in the rocks. Its origin lies in a company that was important in the early history of semiconductors but is rarely mentioned today: Mostek.

Mostek was founded in 1969, and its founder came from Texas Instruments. In an era when DRAM was still at 4K and 16K, Mostek once held a high market share and was an important player in the early DRAM industry. Three of Micron's four founders were former Mostek employees.

In 1978, Micron was founded in the basement of a dental clinic in Boise. At the time, Micron had no manufacturing business at all; its entire business was providing semiconductor design consulting to Mostek.

Far from Silicon Valley, Micron did not follow the script familiar to us today by taking VC (venture capital) money, but instead took money from several local "country gentlemen." This financing method was actually much like early Starbucks. A few successful local businessmen thought these young men were promising, put their heads together, and each invested some money to support them.

But just one year later, this young company suffered a heavy blow. In 1979, after Mostek was acquired by United Technologies, Micron lost its Mostek contract, and its only source of revenue was gone.

After losing its only customer, Micron decided to make chips itself. Building a factory required more money. Through the introductions of those local "country gentlemen," Micron's founders met potato tycoon JR Simplot.

In 1980, JR Simplot was already in his seventies, having grown potatoes all his life, and his company was McDonald's largest supplier of frozen french fries. Without the complex due diligence of VCs and drawn-out investment committee meetings, after visiting Micron's office in the dental clinic basement, JR Simplot invested $1 million and acquired 40% of Micron.

Why would an old man who grew potatoes and sold french fries, already successful and not short of money, invest in a tiny unknown company still in a basement with only four people, and one doing memory chips that had nothing to do with french fries? Perhaps because both fries and chips are called "chips" in English?

Of course, Simplot's $1 million was far from enough to build a factory, but it was equivalent to giving Micron a "credit pass." With his endorsement, Micron successfully obtained loans from local banks and raised money from other Idaho businessmen one after another.

But at the time, the construction cost of a DRAM factory had already reached hundreds of millions of dollars, and Micron could not afford it even if it emptied its pockets.

To keep costs as low as possible, they used the building plans of a supermarket to renovate the plant, expanded the original meat refrigeration space into a clean room, and at the same time purchased large quantities of second-hand equipment discarded by other manufacturers. Eventually, this 50,000-square-foot wafer fab was completed in 1981, costing only about $7 million.

In 1981, Micron began producing 64K DRAM, officially transforming from a chip design company into a true semiconductor manufacturer. And Micron's timing for building the factory can be said to have been extremely exquisite.

In 1981, as the IBM PC began to enter the market, the personal computer industry rose rapidly thereafter. As an indispensable core memory device for computers, demand for DRAM rose rapidly along with it. Micron quickly broke into the market with a smaller, more advanced 64K DRAM design and relatively lower manufacturing costs.

Just three years later, on June 1, 1984, Micron officially listed on Nasdaq. Adjusted for today's stock split basis, its first-day opening price was about $1.35.

Although South Korea now dominates global memory chips, in the early days of the memory industry it was actually American companies that ruled. Intel's first mass-produced product was a memory chip, and well-known American semiconductor companies such as IBM, Texas Instruments, AMD, Motorola, and National Semiconductor had all once been involved in the memory industry.

This pattern persisted until the 1980s, when competition in memory chips entered a different set of rules.

Under these rules, national will became the driving force behind industrial competition. Some countries used administrative power to provide targeted support to specific companies across multiple dimensions such as R&D, funding, and talent. Companies could ignore short-term profits and stock price performance, maintain extremely high-intensity capital expenditure, and even invest counter-cyclically during industry downturns, all in order to outlast their rivals.

From the rise of Japanese companies in the 1980s to the later surge of South Korean companies, both followed this path, and even the specific tactics were much the same: I don't care whether I make money; what I want is for you not to survive.

So Japanese companies used it to crush American companies, and South Korean companies later used it to crush Japanese companies. Under such brutal competition, almost all American semiconductor companies exited the memory business, leaving Micron as the sole survivor.

Yang Yige

Head of the Technology Group, Asset Securitization Research Department, Sunshine Insurance

Memory is very similar to foundry: it is a highly capital-intensive industry that requires large-scale capacity manufacturing and capital expenditure to keep fabs operating. When the memory cycle turns down, current profits cannot cover depreciation on heavy assets, and many companies go bankrupt at this stage.

The core technological barriers in the most critical technical segments remain concentrated in the United States and Europe, some upstream materials and components are concentrated in Japan, and Taiwan and South Korea benefit from lower manufacturing costs and early government support.

But under such a market landscape, why did only Micron survive? After all, the United States has no state-controlled banks that can transfuse cash to Micron, and American capital markets could not tolerate any company investing while losing money.

Here, we have summarized a conclusion and will try to answer the question above: Micron possesses a kind of "American-East Asian duality," combining both Americanness and East Asianness, and it combines the strengths of both sides particularly well.

▍Micron's "Americanness"

Micron is first and foremost an American company, and this is important for understanding Micron's success. Micron's Americanness is first reflected in the fact that it can rely on the United States' developed capital markets and flexible financial instruments. During industry troughs, by leveraging its identity as an American company, it uses extremely shrewd capital operations to expand its own capacity at low cost.

Consider an early case: in 1998, Micron bought Texas Instruments' memory business. The details of the deal are very interesting. At first glance, it was a transaction worth more than $800 million. But in fact, Micron did not pay a single cent in cash. What Micron paid was more than $800 million worth of shares issued in a private placement and $950 million in corporate bonds.

Not only did Texas Instruments not receive cash, it also provided Micron with $750 million in financing to convert the acquired plant to Micron's production processes. Yes, you read that right: it was the seller, Texas Instruments, that put up another $750 million in cash for Micron.

1998 was a cyclical low point for the memory industry, and Texas Instruments was eager to shed its heavily loss-making memory business. Thus Micron, without spending a single cent in cash, took over all of Texas Instruments' memory business and conveniently obtained cross-licenses for a large number of patents.

Beyond sophisticated capital operations, Micron's "Americanness" has another often-overlooked dimension: the law. Starting in the 1980s, Micron repeatedly turned the U.S. trade and judicial systems into competitive tools of its own.

In 1985, when Japanese manufacturers were routing their American peers in the DRAM market, Micron filed an anti-dumping complaint with the U.S. government, ultimately pushing the United States to launch an investigation into Japanese DRAM and helping bring about the 1986 U.S.-Japan Semiconductor Trade Agreement.

In the 1990s, facing the rise of South Korean manufacturers, Micron again initiated anti-dumping investigations. In 2002, after its failed acquisition of Hynix, Micron further pushed for an investigation into South Korean DRAM subsidies, and Hynix ultimately was hit with high anti-subsidy tariffs.

But Micron's most classic "legal maneuver" was still the DRAM price-fixing case.

In 2002, when the U.S. Department of Justice was investigating multiple DRAM manufacturers for colluding to raise prices, Micron was the first to report to the DOJ and apply for leniency, becoming the first company involved in the case to actively cooperate with the investigation, providing key evidence on how the cartel operated and continuing to cooperate with the investigation.

In the end, peers such as Samsung, Hynix, Infineon, and Elpida were hit with huge fines, while Micron, as the whistleblower under the leniency policy, was not criminally prosecuted and did not even pay any fine.

So what Micron's American identity brought it, in addition to sophisticated capital operations, was a set of American laws and regulatory systems that it was naturally familiar with and could leverage, helping it change the rules of the game when competition was at its fiercest.

But other American companies had these two advantages as well, so why did they all exit the market? This actually has to do with the particular nature of Micron's business. At the time, the memory chip players in the American market, such as Intel, IBM, Texas Instruments, and others, all had more profitable businesses, and memory was just one branch of their operations.

Yang Yige

Head of the Technology Team, Asset Securitization Research Department, Sunshine Insurance

Companies like IBM, Texas Instruments, or Intel essentially had other businesses they could develop. For example, Texas Instruments could make analog chips, Intel could make CPUs, and IBM was an ODM computer manufacturer. These companies had other core businesses to support them, so after that round of industry consolidation, they sold off their memory businesses as needed and handed them over to others as needed.

But Micron was different. It had no other business; from the day it was founded, it focused solely on memory. The more the memory industry slumped, the less room it had to retreat, and the more it had to find a way to weather the storm. And the crises of these local giants were seen by it as opportunities to turn the tide.

Later, Micron brought this approach to the East Asian market as well.

▍Micron's "East Asianness"

Starting in 2008, Micron completed several extremely shrewd, jaw-dropping deals that directly produced today's industry reshuffle into the Big Three memory makers. It was also at this stage that, while maintaining its own Americanness, Micron gradually began to East Asianize on the production side.

Qimonda was a company spun off from Germany's Infineon that focused on memory business, and was the "hope of the whole village" for Europe's memory industry at the time. After the 2008 financial crisis, memory prices plummeted, and Qimonda's cash flow became extremely tight. Just as Qimonda was struggling on the edge of life and death, Micron arrived.

Inotera was a company jointly established by Taiwan's Nanya Technology and Infineon, focused on DRAM manufacturing. After Qimonda was spun off from Infineon, Inotera's equity was also transferred to Qimonda. In 2008, Qimonda held about 35% of Inotera's equity.

This time, Micron did not play the share-swap trick again. It readily acquired all of Qimonda's Inotera equity for about $400 million in cash.

Of this $400 million in cash, Inotera itself lent Micron $85 million, and Nanya Plastics, the parent company of Nanya Technology, lent Micron $200 million. Micron's own actual cash outlay was only about $113 million.

In other words, Micron used only about $113 million of its own cash to obtain about 35% of Inotera's equity. What it also pocketed was about 50% of Inotera's capacity allocation rights and the right to introduce Micron's process into Inotera.

Who would pay the cost of introducing Micron's process? Micron certainly would not.

A year later, in August 2009, Inotera raised about $310 million through a share issuance to convert the plant to Micron's process. By then the share price had also recovered, and Micron recognized tens of millions of dollars in accounting gains.

But Inotera was destined to be devoured completely by Micron. In December 2015, Micron acquired the remaining Inotera shares for about $4.1 billion. Of that $4.1 billion, about $2.5 billion was provided by a syndicate of local banks as loans, with collateral including Inotera's plant, land, and some equipment, and about $1 billion was in shares privately placed to the seller, Nanya Technology. So all told, Micron spent only $600 million of its own cash on its books to fully acquire the remaining Inotera shares.

After the acquisition, Inotera was officially renamed "Micron Technology Taiwan" (MTTW).

Then came 2012, which was also a highlight moment for Micron before the AI boom. That year, Japan's "national memory team" Elpida filed for bankruptcy protection, the largest manufacturing bankruptcy in Japanese history.

Without any surprise, Micron showed up again. Micron obtained 100% of Elpida's equity for 200 billion yen, about $2.5 billion at the exchange rate at the time.

But would Micron honestly put up 200 billion yen in cash? Of course not. When the deal was reached, Micron put up only 60 billion yen in cash, about $750 million at the time. The remaining 140 billion yen, about $1.75 billion, was arranged to be paid in installments over six years after the deal closed. The specific arrangement was: 20 billion yen per year from 2014 to 2017, and 30 billion yen per year from 2018 to 2019.

140 billion yen, paid off over six years—harsh enough terms, right? But that was not all. The specific payment method for the 140 billion yen was that after Elpida became a wholly owned subsidiary of Micron, it would continue producing DRAM, and Micron would pay Elpida wafer manufacturing fees, R&D service fees, and management service fees. These fees were denominated in U.S. dollars, and Elpida would use the cash flow generated by continued production to pay creditors, denominated in yen.

In addition, Micron continued to "earn" money from exchange rates.

The deal was reached in July 2012, when Micron paid only a very small deposit. By the time the closing was completed, it was already August 2013. During this period, the yen-to-dollar exchange rate fell from about 80 to 1 to about 98 to 1. In other words, by the actual closing date, the 60 billion yen in cash Micron paid had already dropped from about $750 million to about $615 million. And over the six years of subsequent installment payments, the yen-to-dollar exchange rate fell further, so the final amount paid was far below $2.5 billion.

But saving so much money wasn't actually the main point; more importantly, through the acquisition of Elpida it also gained Taiwan's Rexchip. At the time of its bankruptcy, Elpida held about 65% of Rexchip Electronics (Rexchip), a DRAM manufacturing joint venture between Elpida and Powerchip Technology, and an important production base for Elpida.

After acquiring Elpida, Micron inherited the roughly 65% stake in Rexchip held by Elpida, and separately signed an agreement with Powerchip and its affiliates to acquire 24% of Rexchip. This separate acquisition was worth about NT$10 billion, or roughly US$334 million. Very unusually, this US$334 million was paid directly by Micron in cash.

When the deal closed in 2013, Micron held 24% + 65%, about 89% of Rexchip in total. The following year, in 2014, Micron bought another 9.9% of Rexchip from Kingston for about US$146 million, raising its stake to over 99%. And Rexchip is today's Micron Memory Taiwan (MMT).

Yang Yige

Head of the Technology Group, Asset Securitization Research Department, Sunshine Insurance

When Elpida went bankrupt, its debt pressure was enormous. It had long faced low yields and slow capacity ramp-up, and happened to be hit by an industry downturn. The discounted value of its payments was also clearly below face value.

For the Japanese government and creditor banks, only a large company as buyer could have the corresponding operational capability and be able to take on most of the employees and debt. For Micron, the acquisition brought more technology, capacity and employees. Micron was essentially buying, at a low point in the industry cycle and at a very cheap equity price, a call option on a future industry oligopoly.

Looking back now, it is entirely fair to say that the acquisitions of Elpida and Rexchip were the most important acquisitions in Micron's history, and the cornerstone of Micron becoming today's trillion-dollar giant.

After acquiring Elpida's Hiroshima DRAM fab, Akita packaging and testing plant, and Rexchip's Taichung DRAM fab, Micron's total capacity at the time went from fourth in the world to second, behind only Samsung.

Strategically, Micron had previously been stronger in PC and server DRAM, but Elpida had strong advantages in mobile phones, tablets and low-power DRAM. So the acquisition completed Micron's mobile product line, and Micron later became a DRAM supplier to Apple. Micron spent only a small amount of its own cash to get all of this.

Of course, Micron has not been without its failures. In our episode on Hynix, we also mentioned that Micron once nearly acquired Hynix. But because its offer was too harsh, although the creditor banks agreed, it was unanimously rejected by all of Hynix's directors.

After rejecting Micron's acquisition, Hynix held on for another 10 years and finally got acquired by SK Group. It then boldly bet on HBM and became a big winner in the AI era.

Imagine that in another parallel universe, if Micron had successfully acquired Hynix, today's memory landscape would not be a trio of giants; it would most likely be a duopoly of Micron and Samsung. And without an independent Hynix betting on HBM early, if Nvidia's H100 had not launched on schedule or had performed poorly, today's AI boom might have been delayed.

Having completed these acquisitions, we can no longer view Micron simply as an American company. Because from the perspective of core wafer manufacturing, Micron can almost be seen as an East Asian company.

As of now, Micron has only one plant on US soil, in Manassas, Virginia, which accounts for only about 2% of global memory supply. The vast majority of its other capacity is located in Asian bases. Among them, the core manufacturing and supporting bases for advanced DRAM and HBM are concentrated in four cities: Taoyuan, Taichung, Tongluo in Miaoli, and Tainan.

Taoyuan is one of Micron's earliest DRAM bases, centered on Fab 11. The newly built Fab 11E will introduce EUV and is scheduled to start production in 2028.

Taichung is the largest, handling both advanced DRAM wafer manufacturing and serving as a major HBM packaging and testing hub. There are currently two advanced packaging plants locally, with the second opening in October 2026.

Tongluo in Miaoli is the newest DRAM base expansion. In 2026, Micron bought Powerchip's P5 plant for US$1.8 billion and is building a second wafer fab.

Tainan is a wafer testing base, converted from an old AUO plant, specifically to support DRAM capacity expansion in Taoyuan and Taichung; the first batch of equipment moved in in September 2026.

Japan centers on Hiroshima Fab 15, which handles advanced DRAM wafer manufacturing. This plant came from Elpida, which Micron acquired in 2013, and is currently being upgraded to support future DRAM nodes and AI memory production.

Singapore is Micron's most important NAND manufacturing base; about 98% of Micron's NAND capacity currently comes from here. As memory demand surges in the AI era, Micron is also expanding on two fronts in Singapore: a new NAND fab of about US$24 billion broke ground in early 2026 and is scheduled to start production in the second half of 2028. Meanwhile, an HBM advanced packaging plant of about US$7 billion is also under construction and is expected to contribute HBM capacity from 2027.

Yang Yige

Head of the Technology Group, Asset Securitization Research Department, Sunshine Insurance

Micron is politically an American company and operationally an East Asian company. From an operational perspective, no matter which company, building plants in the US certainly offers no great advantage.

Take TSMC's plant construction in the US as an example: its gross margin will be diluted by two to three percentage points. For TSMC, ramping production at the US plant and employees' willingness to work overtime are also things that cannot be guaranteed. I think the same is true for Micron, except that, unlike TSMC, it is politically an American company.

Micron's East Asian character also explains a question in our minds: with so many problems in American manufacturing, in semiconductor manufacturing—the industry that places the greatest emphasis on efficiency and relies most on high-quality engineers—how has Micron, as an American company, managed to get manufacturing right?

Let us first look at where Micron currently stands in the overall memory industry.

▍Micron's Current Business Situation

If we look purely at company size, there are in fact only three companies worldwide that truly deserve to be called memory giants: Samsung, SK Hynix, and Micron.

In terms of revenue scale and production capacity, Micron is the smallest of the three. Taking the financial data for the second quarter of 2026 as an example, SK Hynix's revenue was about US$55 billion, Micron's revenue in the same period was about US$41.46 billion, and Samsung's semiconductor business revenue was about US$87.7 billion.

In terms of production capacity, the gap is even larger. According to SemiAnalysis estimates, by the end of 2026, Samsung's monthly DRAM capacity will be about 720,000 wafers, SK Hynix's about 595,000 wafers, and Micron's about 385,000 wafers. In the most critical area, HBM, as of the end of 2025, the monthly capacity of both SK Hynix and Samsung was around 160,000 to 170,000 wafers, while Micron had only about 60,000 wafers, roughly one-third of the other two.

However, judging by the growth rate of recent quarters, Micron is the fastest-growing and most profitable of the three.

The latest quarterly financial report shows that Micron's revenue has reached US$54.2 billion, growing another 31% on a high base. For the full fiscal year 2026, revenue was US$133.2 billion, up 256% year on year. Its operating margin also rose from about 25% in early 2025 to more than 80%, and in the second quarter of 2026 it overtook SK Hynix to become the highest of the three.

Why has Micron's explosive momentum been so strong recently? Behind this is actually its business structure.

Samsung's business is relatively complex, burdened by foundry and chip design operations with relatively low profit margins, so we will not compare it here. But Micron's and SK Hynix's businesses are actually highly similar, and the two can be looked at together.

Both Micron's and SK Hynix's businesses mainly consist of two major blocks, DRAM and NAND, and the proportions are very similar: DRAM accounts for nearly 80% and NAND about 20%. Within DRAM, there are also two lines: one is traditional DRAM, mainly general-purpose memory such as DDR5 for servers and LPDDR for phones and tablets. The other is HBM, which is also the most important reason for the difference.

SK Hynix began betting on HBM very early, staking almost its entire fortune on it. As early as 2013, it jointly launched HBM1 with AMD, then advanced generation by generation, freeing up more and more DRAM capacity specifically for HBM. By now, about 30% of its DRAM capacity is used for HBM, and by 2027 that proportion is expected to rise to nearly 40%.

This determination to bet early and invest heavily has earned it a near-monopoly leading position in HBM today: in the global HBM market, SK Hynix alone takes about 60% of the share and is the undisputed king.

By contrast, Micron started very late. It did not launch its first HBM product until 2020, by which time SK Hynix and Samsung had already delivered two generations of products, HBM and HBM2. As a result, HBM now accounts for only about 20% of Micron's DRAM revenue, while HBM accounts for nearly half of SK Hynix's DRAM revenue.

Herein lies the most critical and most counterintuitive change in this cycle: the product whose price has risen the most is actually not HBM, but the neglected traditional DRAM.

Because all three giants are busy shifting capacity to higher-profit HBM, a large chunk of supply of the general-purpose memory used in servers, PCs, and phones has been siphoned away, leading to a full-scale shortage, and prices have soared.

Data shows that in the first quarter of 2026, traditional DRAM contract prices rose more than 90% quarter on quarter, and they continued to rise in the second quarter. By contrast, HBM prices were instead held down in advance, because the leading companies mostly signed early long-term agreements with major customers that locked in prices.

Yang Yige

Head of the Technology Group, Asset Securitization Research Department, Sunshine Insurance

HBM contracts are different from past traditional DRAM and NAND contracts. Traditional DRAM and NAND also have long-term agreements, but prices are usually negotiated quarterly or semiannually and are closer to market rates; only volume is locked, not price, and there are no penalty payments. HBM first emerged to serve AI, and pricing has always been negotiated once a year, mostly around the end of each year.

In other words, this year's HBM prices were set at the end of last year, when the overall memory market had not yet begun its frenzied price increases.

So recently a situation has emerged: SK Hynix sells the most HBM, but its prices are precisely locked in by long-term contracts. When it tilts all its capacity toward HBM, the capacity left for traditional DRAM is naturally squeezed.

Micron's situation is exactly the opposite. In the past, a low HBM share was a shortcoming for which Micron was criticized. But in this cycle, general-purpose DRAM has risen even more sharply than HBM, and Micron's larger exposure in that area has instead allowed it to capture this round of price increases more fully and directly. This "more balanced application structure" is now often overshadowed by the HBM halo, but it is in fact one of Micron's biggest advantages at this stage.

Micron's current business revenue is divided into four segments: Cloud Memory, Core Data Center, Mobile and Client, and Automotive and Embedded. Data for the latest quarter show that revenue from these four businesses was $16.28 billion, $18 billion, $13.11 billion, and $6.82 billion, respectively. Although the two data center-related businesses remain the largest, the mobile, client, automotive, and embedded businesses together still contributed about 37% of revenue, a considerable base.

Especially in automotive and industrial markets, Micron has built up long-term customer relationships and certification barriers. These businesses may not grow as fast as HBM, but they offer stability, high certification barriers, and strong customer stickiness—typical "ballast" businesses.

Rob Li

Managing Partner at Amont Partners, New York

After SK Hynix released its quarterly report, the market judged its operating income to be below expectations, and the stock fell more than 10% that day. A core reason is that HBM accounts for the highest share of SK Hynix's business.

In the past this was a good thing, and the market liked it, because HBM is the most advanced product. But from the fourth quarter of last year to the first half of this year, prices for ordinary memory also rose sharply, while HBM consumes more wafers and costs more than ordinary memory, so its gross margin during this period was actually lower than that of ordinary memory.

In other words, under the same capacity conditions, the more resources allocated to HBM, the more it drags down the company's overall gross margin rather than helping it.

However, the current higher gross margin of traditional memory is only temporary, and HBM remains the key to winning in the future.

▍The Focus of Competition: HBM

Today, HBM is becoming a key variable for the memory industry to ride out cycles. In the past, DRAM and NAND were highly standardized, and memory chips were more like a "commodity," with prices and profits fluctuating sharply with supply and demand cycles, making it hard for manufacturers to maintain stable profitability.

But HBM is different. It requires deep co-design with specific chips for specific customers, must pass custom certification, and each company's stacking and base die solutions differ, making it closer to a non-standardized "custom product." As a result, its technical threshold is higher and customer validation cycles are longer.

As AI computing demand continues to grow, HBM not only brings higher unit prices and profit margins but also gives memory manufacturers stronger order certainty and bargaining power.

So what the market calls the "revaluation of memory value" today is not about the recent price increases in memory chips, but more about how HBM is changing the profit model of this highly cyclical industry, making cycles weaker and smoother.

Yang Yige

Head of the Technology Team, Securities Research Department, Sunshine Insurance Asset Management

If things still follow the old market-based approach, with prices negotiated each quarter based on contract prices, then the higher the share of traditional DRAM and NAND, the greater the cyclical risk. Only when the share of products like HBM—which have long-term agreements and strong demand momentum—is high can cyclical resilience and earnings stability improve.

So Micron is also well aware that it must not fall behind in HBM. In fact, Micron recognized the importance of the "stacked memory" track very early; it just bet on the wrong standard in its early years.

As early as 2011, Micron, together with Samsung, IBM, and others, launched a stacked memory technology called HMC (Hybrid Memory Cube). That was even earlier than SK Hynix's bet on HBM. The idea behind this technology was to stack multiple layers of DRAM dies together with a separate logic control die, install the memory controller directly inside the memory stack, and rely on TSV through-silicon via technology to vertically stack multiple chip layers, forming a three-dimensional cube.

HMC had many technical highlights and was once used in supercomputers, network processors, and other scenarios. But compared with HBM, it consumed more power to transmit the same amount of data and was also hard to adapt to GPUs that are themselves high-power, so it was not adopted by Nvidia or AMD.

In May 2017, former SanDisk co-founder Sanjay Mehrotra became Micron's CEO. After taking office, he decided to stop pursuing HMC and switch to HBM, which had already become the industry standard—widely regarded as one of Micron's most important course corrections in recent years.

But by then, Micron was already a full two generations behind in HBM. Even after Micron launched its first HBM product in 2020, its development pace remained lukewarm. In June 2022, SK Hynix had already begun mass production of HBM3, becoming the world's first manufacturer to enter Nvidia's H100 supply chain with HBM3. But at Micron, there was still no movement for a long time.

What truly turned things around for Micron was a bold decision: it simply skipped an entire generation, HBM3, and bet all its resources on the next generation, HBM3E.

That leap paid off. In February 2024, Micron announced mass production of 24GB HBM3E and squeezed into Nvidia's H200 supply chain, shipping alongside the H200 in the second quarter of that year. From this generation on, Micron's catch-up clearly accelerated: it followed up with higher-capacity 12-Hi HBM3E and kept pushing toward the next-generation HBM4.

By the HBM4 generation, Micron even pulled off a impressive turnaround. In the first quarter of 2026, a full quarter ahead of its original schedule, it began mass production of 36GB, 12-layer stacked HBM4, with bandwidth exceeding 2.8TB per second, pin speeds above 11Gbps, and power consumption more than 20% lower than its own previous generation.

In June of the same year, Nvidia CEO Jensen Huang publicly confirmed that all three—SK Hynix, Samsung, and Micron—had passed HBM4 certification and entered mass production, jointly supplying the next-generation Vera Rubin platform. This was also the first time in HBM history that a new-generation product had all three major suppliers from the start. At this point, Micron finally shed its "latecomer" label and, in terms of product specifications, once again stood on par with Samsung and SK Hynix in the same generation.

However, "being able to make it" and "how much you can make" are still two different things. According to estimates by supply chain analysts, in the order allocation for this generation of Vera Rubin, SK Hynix alone took 60% to 70%, Samsung took 20% to 30%, and what was left for Micron was only a small remaining slice. In terms of revenue share, Micron accounts for only 20% of the HBM market, still a clear gap behind SK Hynix.

Rob Li

Managing Partner at Amont Partners, New York

At the current stage, the Pin Speed (per-pin transfer rate) of the several manufacturers all meets specifications. From Micron's own perspective, it does not have a particularly strong incentive, because its capacity is the smallest of the three. Micron has said repeatedly that as long as its HBM market share can remain consistent with its share of DRAM capacity, that is acceptable. From Nvidia's perspective as a customer, it will use all three suppliers; what it needs most is for them to guarantee supply volume.

So for Micron, beyond continuing to catch up technologically, the key hurdle to cross in the future is still the "share" firmly held by SK Hynix through its early binding and several-fold capacity advantage. However, HBM competition continues. The next generation, HBM4E, is becoming the next battleground for the three giants.

Micron plans to introduce the most advanced EUV lithography technology into mass production for the first time with HBM4E. But our guest believes that shifting to EUV technology does not mean it gives Micron an opportunity to overtake on a curve technologically. Judging from the current landscape, SK Hynix still firmly occupies the top spot in HBM.

Yang Yige

Head of the Technology Group, Securities Research Department, Sunshine Insurance Asset Management

When process pressure keeps growing, it becomes very difficult not to use EUV. Memory is an industry based on technology and won by capacity scale. Once everyone adopts EUV, you can only say the technology generations have been leveled. For HBM, packaging capability also matters—how to build the stack more stably and better. At present, I think SK Hynix is better, while Samsung and Micron are roughly similar.

There is another important point: if we look carefully at the timing of Micron's accelerated HBM catch-up, you may notice that it happened exactly against the broader backdrop of the United States revitalizing chips and manufacturing, which has also changed Micron's identity today.

HBM is one of the most critical infrastructure pieces determining AI chip performance today, and among the manufacturers globally capable of large-scale production of advanced HBM, Micron is the only American company. This means that if the United States wants to control the complete AI supply chain from GPUs and advanced packaging to HBM, it cannot rely only on chip design companies such as Nvidia and AMD; it must also keep Micron. So the United States began looking for ways to "pull" Micron back from Asia.

In 2022, the United States passed the CHIPS and Science Act, placing advanced semiconductor manufacturing at the top of national industrial policy. As the only domestic memory leader in the United States, Micron became the most central subsidy target.

The conditions offered by the U.S. government were highly attractive, including direct subsidies, loans, tax incentives, plus land and infrastructure support provided by state governments.

Micron's response was a large-scale plan to build factories in the United States. According to data on Micron's official website, it plans to invest more than $250 billion in total in the United States by 2035.

These investments mainly fall in three places:

The first site is Micron's hometown, Boise, Idaho. In 2022, Micron announced an investment of about $25 billion to build an advanced DRAM wafer fab next to its headquarters—the first new advanced memory factory built in the United States in decades. In 2025, Micron announced it would build a second factory. The two factories will be directly integrated with Micron's existing Idaho R&D center and will also undertake R&D and manufacturing of advanced products including HBM in the future.

The second site is Clay, New York, where the scale is even more exaggerated. Micron plans to gradually build a mega-factory campus consisting of up to four advanced DRAM fabs over more than 20 years, with total investment of up to $100 billion. At present, New York's first fab has already broken ground in 2026.

In addition, the long-established Manassas, Virginia factory has also been included in expansion and upgrading. It does not pursue the most advanced process, but mainly brings mature 1α DRAM back to the United States, dedicated to long-cycle markets such as automotive, defense, and industrial.

If you look at these factories together, what Micron is trying to do is put the most advanced DRAM manufacturing capability back in the United States. With Boise as the core for R&D and advanced DRAM manufacturing, taking on the most cutting-edge processes and HBM-related capabilities, Clay, New York as the largest future advanced DRAM production base, and Manassas, Virginia taking on mature processes to serve markets such as automotive, industrial, and defense that have stable demand for mature nodes.

Micron said it plans to place about 40% of its DRAM production capacity in the United States. Correspondingly, the U.S. share of global advanced memory manufacturing will rise from less than 2% today to about 10%.

However, this is still a marathon measured in decades. The fastest-moving Boise Phase 1 will not begin production until around 2027. The largest factory cluster in Clay, New York, is not even scheduled until the 2030s or later.

Real obstacles have already emerged. The New York campus has repeatedly come under pressure due to labor shortages and long construction cycles. Micron has had to shift funds between states and prioritize projects that can ship sooner.

So bringing advanced production capacity back to the United States has never been something that a single bill or a single subsidy can immediately deliver.

At least in the first half of this decade, when we say Micron is "America's lone memory player," most of the advanced capacity it prides itself on is, in fact, still on the other side of the Pacific.

Rob Li

Managing Partner of Amont Partners, New York

When the CEO says he hopes to put 40% of production capacity in the United States in the future, note that he did not mention a timeline, nor say whether it would be next year, three years from now, ten years from now, or twenty years from now. So there is a lot of room for interpretation in between, and he can make that statement work either way.

From Micron's perspective, if it expands production capacity in the United States, the most recent news points to around 2030. Of course, another decade after that, it can continue to expand gradually, but I think the scale that can actually be implemented will still be relatively limited.

Not long ago, a heated "battle" broke out between Apple and Micron. As memory chip prices surged, Apple had to raise prices on some products and directly pinned the cost pressure on Micron. To find cheaper supply, Apple sought support from the U.S. government, hoping to expand the supply of Chinese memory chips.

But when Micron heard about this, it not only publicly pushed back at Apple but also pressured the U.S. government in turn, trying to block the plan. What is interesting about this conflict is not actually who wins between Apple and Micron, but that Micron's position has completely changed.

In the past, downstream giants like Apple often had stronger bargaining power. Suppliers had to revolve around Apple, finding ways to cut costs and secure supply.

But in the AI era, when memory has become one of the scarcest resources for AI computing power, Micron has leapfrogged into the position of holding the power of discourse. From the basement of a dental clinic in Idaho to becoming one of only three DRAM giants in the world today. From struggling to survive industrial wars in Japan and South Korea to surviving again and again by using capital markets and trade rules. And now, the U.S. government is willing to put up billions of dollars just to pull its advanced manufacturing capability back to the United States.

And all of this may be the reward Micron has finally earned after nearly 50 years of betting on memory and enduring cycle after cycle.

The views in this article represent only the author's own. The 36Kr platform only provides information storage space services.+144

Good article, needs your encouragementSilicon Valley 101Guest Author12Favorite+10CommentOpen WeChat "Scan", open the webpage and click the share button in the upper right cornerWeiboImmersive readingBack to topReportJoin the commentsTens of millions of comments, be friendly firstLog in to join the discussionSubmit comment0/1000You may also like these articlesFrom reporter to billionaire, the "greatest liberal arts student" bet on AIBreaking news! After OpenAI, AI once again cracks the generalized Riemann hypothesisNot yet IPO'd, 600 OpenAI employees have already cashed out $6.6 billionCrayfish has gone cold, why is Muse popular? Was it reasonable to poach people at sky-high prices back then? Alexander Wang's latest interview responds to everythingJust now, Anthropic announced that the singularity has arrived! AI may push U.S. GDP to $44 trillionWishful programming released, who says Liang Wenfeng doesn't spoil "newbies"?The grandson of the world's richest man is about to go bankrupt after 30 years of entrepreneurshipThe $4.99 AI doctorFormer Huawei Car BU world model foundation model head starts a business, Hangzhou embodied brain company receives nearly 100 million yuan in financing | Hardcore exclusiveSilicon Valley 101Guest Author

"Silicon Valley 101" is a tech podcast sharing the freshest current technology and ideasPublish articles235 articles

Recent contentBehind Micron's rise: the survival battle, capital game, and AI rebirth of America's "lone memory player"5 hours agoSuper El Niño is here, will our daily necessities really become more expensive because of it?2026-10-09Who sets questions, sells questions, and grades papers for large models? Let's talk about the wild growth of the AI data industry2026-10-08Read more content, poke here hardChoose reliable AI, see real reviewsView 36Kr AI reviewsOfficial communication communityJoinConsultationProject review and onboarding contact 36KrProject recommendationsSubscription accountFollow

Next articleThe president of a 60 billion company, 59 years old, is going to take over the "shift" of the son of a UP host

A father who took professional managers to the extreme, and a son who built a business from scratch, appear in the same frame on a business registration form.

36Kr Europe36Kr Europe36Kr EuropeAI Product DailyOnline rumor information report entranceHot recommendations

Push and interpret cutting-edge, substantive tech startup and investment news

Primary market financial information and system service provider

Focus on outstanding entrepreneurs worldwide, with a project financing rate close to 97%, leading the industry